There's an art to subtitling that goes beyond mere speech-to-text processing. Sometimes it's better to paraphrase dialog to reduce the amount of text that needs to be read. Sometimes you need to name a voice as unknown, to avoid spoilers. Sometimes the positioning on the screen matters. I hope the model can be made to understand all this.
They're using Whisper for speech to text, and some other small model for basic translation where necessary. It will not do speaker identification (diarization), and certainly isn't going to probe into narrative plot points to figure out if naming a character is a reveal. It isn't going to place text on the screen according to the speaker's frame place, nor for least intrusion. It's just going to have a fixed area where a best effort at speech to text is performed, as a last resort where the alternative is nothing.
Obviously it would be preferred to have carefully crafted subtitles from the content creator, translating if the desired language isn't available but still using all the cues and positions. Secondly to have some carefully crafted community subtitles from opensubtitles or the like, maybe where someone used "AI" and then hand positioned/corrected/updated. Failing all that, you fall to this.